Cooperative Recurrent Neural Networks For Multi-category Classification And Speech Enhancement

Multi-category classification and speech enhancement has been widelyused in the fields of scientific and engineering applications and is twopopular issues in world. Their solution algorithms have been greatlydeveloped, but algorithm complexity and estimation problems still needto be studiedd. The first research part of this thesis studied themulti-category classification. By using Support Vector Machinetechnology, we first decompose the Multi-category classification probleminto the integration of two constrained optimization problems, and thenpropose a cooperative recurrent neural network classifiers. Thisclassifiers is composed of one neural network for bias estimator andanother neural network for support vector learning. In the second researchpart of this thesis,the problem of speech enhancement is reformulated asboth autoregressive parameter estimation and speech signal prediction.We then propose a cooperative recurrent neural network for thespeech enhancement. It automatically combines an adaptive recursiveneural networks and Kalman filters recursive network. Comparison withrelated algorithms, including a single recurrent neural network, theproposed cooperative neural network has a lower computational complexity,and ensures global convergence to the optimal solution. Simulationresults show that cooperative recurrent neural network can be effectivelyapplied to Multi-category classification problem and Speech enhancementproblem. Moreover, it can obtain better estimate than relatedalgorithms.